WEBVTT - Deepfakes and the Future of Truth 0:00:15.250 --> 0:00:26.810 Pushkin, you're listening to Brave New Planet, a podcast about 0:00:26.850 --> 0:00:31.010 amazing new technologies that could dramatically improve our world. Or 0:00:31.410 --> 0:00:33.970 if we don't make wise choices, could leave us a 0:00:33.970 --> 0:00:44.690 lot worse off. Utopia or dystopia. It's up to us 0:00:45.810 --> 0:00:52.810 pash the eye. On July sixteenth, nineteen sixty nine, Apollo 0:00:52.850 --> 0:00:58.290 eleven blasted off from the Kennedy Space Center near Cape Canaveral, Florida. 0:00:58.650 --> 0:01:03.290 Twenty five million Americans watched on television as the spacecraft 0:01:03.410 --> 0:01:08.890 ascended toward the heavens, carrying commander Neil Armstrong, Lunar Module 0:01:08.930 --> 0:01:14.410 pilot Buzz Aldron, and Command Module pilot Michael Collins their 0:01:14.410 --> 0:01:18.330 mission to be the first humans in history to set 0:01:18.410 --> 0:01:23.810 foot on the Moon. Four days later, on Sunday, July twentieth, 0:01:24.210 --> 0:01:27.970 the lunar module separated from the command ship and soon 0:01:28.210 --> 0:01:34.490 fired its rockets to begin its lunar descent. Five minutes later, 0:01:35.330 --> 0:01:40.250 disaster struck about a mile above the Moon's surface. Program 0:01:40.290 --> 0:01:43.890 alarms twelve O one and twelve O two sounded loudly, 0:01:43.930 --> 0:01:49.810 indicating that the mission computer was overloaded, and then, well, 0:01:51.130 --> 0:01:59.530 every American knows what happened next. Lost date of five 0:02:05.770 --> 0:02:10.130 good evening, my fellow Americans, President Richard Nick addressed a 0:02:10.250 --> 0:02:14.970 grieving nation. Fades has ordained that the men who went 0:02:15.010 --> 0:02:19.010 to the Moon to explore in peace will stay on 0:02:19.050 --> 0:02:25.930 the Moon to rest in peace. These brave men, Neil 0:02:26.250 --> 0:02:32.130 Armstrong and Edwin Auburn, know that there's no hope for 0:02:32.210 --> 0:02:37.290 their recovery, but they also know that there is hope 0:02:37.410 --> 0:02:42.450 for mankind in their sacrifice. He ended with the now 0:02:42.570 --> 0:02:46.370 famous words for every human being who looks up at 0:02:46.370 --> 0:02:49.530 the Moon and the nights to come, will know that 0:02:49.530 --> 0:02:54.610 there is some corner another word that is forever mankind. 0:02:58.130 --> 0:03:02.130 Wait a minute, that never happened. The Moon mission was 0:03:02.170 --> 0:03:07.170 a historic success. The three astronauts returned safely to ticker 0:03:07.210 --> 0:03:11.130 tape parades and a celebrity thirty eight day world tour. 0:03:11.930 --> 0:03:15.730 Those alarms actually did sound, but they turned out to 0:03:15.730 --> 0:03:21.130 be harmless. Nixon never delivered that speech. His speechwriter had 0:03:21.130 --> 0:03:24.450 written it, but it sat in a folder labeled an 0:03:24.530 --> 0:03:30.650 event of Moon disaster until now. The Nixon you just 0:03:30.810 --> 0:03:34.050 heard is a deep fake, part of a seven minute 0:03:34.090 --> 0:03:39.650 film created by artificial intelligence deep learning algorithms. The fake 0:03:39.810 --> 0:03:43.090 was made by the Center for Advanced Virtuality at the 0:03:43.090 --> 0:03:46.930 Massachusetts Institute of Technology as part of an art exhibit 0:03:47.210 --> 0:03:50.930 to raise awareness about the power of synthesized media. Not 0:03:51.050 --> 0:03:54.570 long ago, something like this would have taken a lot 0:03:54.730 --> 0:03:58.410 of time and money, But now it's getting easy. You 0:03:58.450 --> 0:04:01.450 can make new paintings in the style of French Impressionism, 0:04:01.770 --> 0:04:06.330 revived dead movie stars, help patients with nor degenerative disease, 0:04:06.810 --> 0:04:09.570 or soon maybe take a class on a tour of 0:04:09.610 --> 0:04:14.090 ancient rome. But as the technology quickly becomes democratized, we're 0:04:14.130 --> 0:04:17.690 getting to the point where almost anyone can create a 0:04:17.730 --> 0:04:21.210 fake video of a friend, an ex lover, a stranger, 0:04:21.650 --> 0:04:26.650 or a public figure that's embarrassing, pornographic, or perhaps capable 0:04:26.690 --> 0:04:31.450 of causing international chaos. Some argue that in a culture 0:04:31.490 --> 0:04:36.290 where fake news spreads like wildfire and political leaders deny 0:04:36.410 --> 0:04:40.530 the veracity of hard facts, deep fake media may do 0:04:40.570 --> 0:04:47.090 a lot more harm than good. Today's big question will 0:04:47.170 --> 0:04:52.570 synthesized media unleash a new wave of creativity or will 0:04:52.570 --> 0:04:57.170 it erode the already tenuous role of truth in our democracy? 0:04:57.290 --> 0:05:00.850 And is there anything we can do to keep it 0:05:00.890 --> 0:05:12.090 in check. My name is Eric Lander. I'm a scientist 0:05:12.090 --> 0:05:15.010 who works on ways to improve human health. I helped 0:05:15.050 --> 0:05:17.770 lead the Human Genome Project, and today I lead the 0:05:17.810 --> 0:05:21.930 Broad Institute of MIT and Harvard. In the twenty first century, 0:05:22.290 --> 0:05:26.770 powerful technologies have been appearing at a breathtaking pace related 0:05:26.810 --> 0:05:31.290 to the Internet, artificial intelligence, genetic engineering, and more. They 0:05:31.330 --> 0:05:35.650 have amazing potential upsides, but we can't ignore the risks 0:05:35.730 --> 0:05:38.770 that come with them. The decisions aren't just up to 0:05:38.890 --> 0:05:42.730 scientists or politicians. Whether we like it or not, we 0:05:43.450 --> 0:05:46.530 all of us are the stewards of a brave New Planet. 0:05:47.090 --> 0:05:50.850 This generation's choices will shape the future as never before. 0:05:53.410 --> 0:05:57.490 Coming up on today's episode of Brave New Planet, I 0:05:57.530 --> 0:06:01.850 speak with some of the leaders behind advances in synthesized media. 0:06:02.050 --> 0:06:05.050 You could, certainly, by the way, generate stories that could 0:06:05.570 --> 0:06:09.370 be fresh and interesting and new and personal for every child. 0:06:09.610 --> 0:06:13.410 We got emails from people who were quadruplegic and they 0:06:13.450 --> 0:06:16.090 asked us if we could make them dance. We hear 0:06:16.170 --> 0:06:19.810 from experts about some of the frightening ways that bad 0:06:19.850 --> 0:06:23.850 actors can use deep fakes. Creditors would chime in and say, 0:06:24.090 --> 0:06:26.970 you can absolutely make a deep fake sex video of 0:06:26.970 --> 0:06:29.770 your ex with thirty pictures. I've done it with twenty. 0:06:30.010 --> 0:06:31.850 Here's the things that keep me up at night right 0:06:32.530 --> 0:06:35.410 a video of Donald Trump saying I've launched nuclear weapons 0:06:35.410 --> 0:06:38.770 against Iran, and before anybody gets around to firing out 0:06:38.810 --> 0:06:40.890 whether this is real or not, we have global nuclear 0:06:40.970 --> 0:06:45.410 outdown and we explore how we might prevent the worst abuses. 0:06:46.250 --> 0:06:52.050 It's important that younger people advocate for the Internet that 0:06:52.090 --> 0:06:54.690 they want. We have to fight for it. We have 0:06:54.770 --> 0:07:04.650 to ask for different things. Stay with us, Chapter one, 0:07:05.210 --> 0:07:10.050 Abraham Lincoln's Head. To begin to understand and the significance 0:07:10.050 --> 0:07:13.530 of deep fake technology, I went to San Francisco to 0:07:13.610 --> 0:07:17.170 speak with a world expert on synthetic media. My name 0:07:17.330 --> 0:07:22.170 is Alexei or sometimes called Alyosha Afros, and I'm a 0:07:22.170 --> 0:07:26.650 professor at UC Berkeley and Computer Science and Lexical Engineering Department. 0:07:27.170 --> 0:07:33.050 My research is on computer vision, computer graphics, machine learning, 0:07:33.770 --> 0:07:39.570 various aspects of artificial intelligence. Where'd you grow up. I 0:07:39.650 --> 0:07:43.530 grew up in Saint Petersburg in Russia. I was one 0:07:43.570 --> 0:07:47.970 of those geeky kids playing around with computers or dreaming 0:07:47.970 --> 0:07:55.170 about computers. My first computer was actually the first Soviet 0:07:55.650 --> 0:07:59.610 personal computer. So you actually are involved in making sort 0:07:59.650 --> 0:08:04.890 of synthetic content, synthetic media, that's right. Alexei has invented 0:08:04.970 --> 0:08:08.970 powerful artificial intelligence tools, but his lab also has one 0:08:09.250 --> 0:08:13.130 full ability to use computers to enhance the human experience. 0:08:13.850 --> 0:08:17.250 I was struck by a remarkable video on YouTube created 0:08:17.290 --> 0:08:21.090 by his team at Berkeley. So this was a project 0:08:21.130 --> 0:08:26.530 that actually was done by my students who didn't even 0:08:26.570 --> 0:08:30.290 think of this as anything but a silly little toy 0:08:30.370 --> 0:08:34.370 project of trying to see if we could get a 0:08:34.490 --> 0:08:39.290 geeky computer science student to move like a ballerina. In 0:08:39.330 --> 0:08:43.050 the video, one of the students, Carolyn cham dances with 0:08:43.130 --> 0:08:46.730 a skill and grace of a professional despite never having 0:08:46.730 --> 0:08:50.890 studied ballet. The idea is, you take a source actor 0:08:50.970 --> 0:08:55.090 like a ballerina. There is a way to detect the 0:08:55.250 --> 0:08:59.770 limbs of the dancer, have a kind of a skeleton extracted, 0:09:00.210 --> 0:09:04.090 and also have my student just move around and do 0:09:04.210 --> 0:09:08.050 some geeky moves. And now we're basically just going to 0:09:08.090 --> 0:09:14.090 try to sympathize the appearance of my student driven by 0:09:14.130 --> 0:09:16.930 the skeleton of the ballerina. Put it all together, and 0:09:16.970 --> 0:09:21.810 then we have our grad student dancing pirouets like a ballerina. 0:09:23.290 --> 0:09:27.850 Through artificial intelligence, Carolyn's body is puppeteered by the dancer. 0:09:28.130 --> 0:09:30.890 We weren't even going to publish it, but we just 0:09:31.370 --> 0:09:35.610 released a video on YouTube called Everybody Dance Now, and 0:09:36.490 --> 0:09:40.650 somehow it really touched the nerve. Well, there's been an 0:09:40.650 --> 0:09:44.890 explosion recently a new ways to manipulate media. Alexei notes 0:09:44.970 --> 0:09:49.410 that the idea itself isn't new, It has a long history. 0:09:49.770 --> 0:09:53.450 I can't help but ask, given that you come from Russia. 0:09:53.770 --> 0:09:58.770 One of the premier users of doctoring photographs I think 0:09:58.930 --> 0:10:03.930 was Stalin, who used the ability to manipulate images for 0:10:04.130 --> 0:10:07.650 political effect. How did they do that? Can you think 0:10:07.650 --> 0:10:10.650 of examples of this and like what was the technology? Then? 0:10:11.370 --> 0:10:17.890 The urge to change photographs has been around basically since 0:10:17.890 --> 0:10:20.930 the invention of photography. For example, there is a photograph 0:10:21.010 --> 0:10:25.650 of Abraham Lincoln that still hangs in many classrooms. That's fake. 0:10:25.730 --> 0:10:30.450 It's actually Calhoun with Lincoln's head attached to it. Alexei's 0:10:30.490 --> 0:10:34.570 referring to John C. Calhoun, the South Carolina senator and 0:10:34.770 --> 0:10:39.970 champion of slavery. A Civil War portrait artist superimposed a 0:10:40.050 --> 0:10:44.530 photo of Lincoln's head onto an engraving of Calhoun's body 0:10:45.090 --> 0:10:49.650 because he thought Lincoln's gangly frame wasn't dignified enough, and 0:10:49.730 --> 0:10:52.570 so they just said Okay, we can use Calhoun. Let's 0:10:52.730 --> 0:10:55.690 slap the Lincoln's head on his body. And then, of course, 0:10:56.250 --> 0:10:59.250 as soon as you go into the twentieth century, as 0:10:59.250 --> 0:11:03.370 soon as you get to dictatorships, this is a wonderful 0:11:04.050 --> 0:11:07.850 toy for a dictator to use. So again, Stalin was 0:11:08.690 --> 0:11:12.370 big fan of this. He would get rid of people 0:11:12.450 --> 0:11:15.850 in photographs once they were out of favor, or once 0:11:15.890 --> 0:11:20.170 they got jailed or killed. He would just basically get 0:11:20.170 --> 0:11:25.810 them scratched out with reasonably crude techniques. Hitler did it, 0:11:25.930 --> 0:11:29.090 Mao did it, Castro did it, Bresnev did it. I'm 0:11:29.090 --> 0:11:32.530 sure US agencies have done it. Also, we have always 0:11:32.570 --> 0:11:36.610 manipulated images with a desire to change history. This is 0:11:36.650 --> 0:11:39.930 Honi f Reed. He's also a professor at Berkeley and 0:11:39.970 --> 0:11:43.130 a friend of Alexey's. I'm a professor of computer science 0:11:43.130 --> 0:11:47.210 and I'm an expert in digital forensics, where Alexei works 0:11:47.250 --> 0:11:50.930 on making synthetic media. Honey has devoted his career to 0:11:51.170 --> 0:11:55.250 identifying when synthetic media is being used to fool people, 0:11:55.810 --> 0:12:00.730 that is, spotting fakes. He regularly collaborates on this mission 0:12:00.730 --> 0:12:05.130 with Alexey so I, met Alyosha efros Ten, twenty years ago. 0:12:05.570 --> 0:12:11.090 He is really incredibly creative and clever guy, and he 0:12:11.690 --> 0:12:14.050 has done what I consider some of the most interesting 0:12:14.090 --> 0:12:16.650 work in computer vision and computer graphics over the last 0:12:16.690 --> 0:12:20.810 two decades. And if you really want to do forensics, well, 0:12:20.930 --> 0:12:23.290 you have to partner with somebody like Aliosha. You have 0:12:23.370 --> 0:12:26.570 to partner with a world class mind who knows how 0:12:26.570 --> 0:12:29.410 to think about the synthesis side so that you can 0:12:29.450 --> 0:12:32.890 synthesize the absolute best content and then think about how 0:12:32.930 --> 0:12:35.370 to detect it. I think it's interesting that if you're 0:12:35.410 --> 0:12:38.250 somebody on the synthesis side and developing the forensic there's 0:12:38.290 --> 0:12:39.890 a little bit of a jekylin hide there, and I 0:12:39.930 --> 0:12:44.290 think it's really fascinating. You know, the idea of altering photos, 0:12:44.930 --> 0:12:47.850 it's not entirely new. How far back does this go? 0:12:48.730 --> 0:12:51.450 So we used to have in the days of Stalin, 0:12:51.970 --> 0:12:57.690 highly talented, highly skilled, time consuming, difficult process of manipulating images, 0:12:58.170 --> 0:13:03.090 removing somebody, erasing something from the image, splicing faces together. 0:13:03.650 --> 0:13:06.810 And then we moved into the digital age where now 0:13:06.850 --> 0:13:10.290 a highly talented digital artist could remove one face and 0:13:10.330 --> 0:13:13.290 add another phase, but it was still a time consuming 0:13:13.330 --> 0:13:17.090 and required scale. In nineteen ninety four, The makers of 0:13:17.090 --> 0:13:21.170 the movie Forrest Gump won an Oscar for Visual Effects 0:13:21.250 --> 0:13:25.410 for their representations of the title character interacting with historical 0:13:25.530 --> 0:13:29.610 figures like President John F. Kennedy gratulating how do the 0:13:29.690 --> 0:13:33.690 field being all Americans? It's very good congratulation. How do 0:13:33.730 --> 0:13:40.090 you feel I got? I believe that he had. Now 0:13:40.450 --> 0:13:42.770 computers are doing all of the heavy lifting of what 0:13:42.930 --> 0:13:46.290 used to be relegated to talented artists. The average person 0:13:46.370 --> 0:13:50.130 now can use sophisticated technology to not just capture the recording, 0:13:50.170 --> 0:13:53.170 but also manipulate it and then distribute it. The tools 0:13:53.330 --> 0:13:57.290 used to create synthetic media have grown by leaps and bounds, 0:13:57.410 --> 0:14:00.330 especially in the past few years, and so now we 0:14:00.410 --> 0:14:04.290 have technology broadly called deep fake, but more specifically should 0:14:04.330 --> 0:14:08.490 be called synthesized content, where you point an image or 0:14:08.530 --> 0:14:11.850 a video or an audio to an AI or machine 0:14:11.930 --> 0:14:15.210 learning system and it will replace the face for you. 0:14:15.450 --> 0:14:16.850 I mean it can do that in an image, it 0:14:16.890 --> 0:14:19.210 can do that in a video, or it can synthesize 0:14:19.250 --> 0:14:25.890 audio for you in a particular person's voice. It's becomes 0:14:25.970 --> 0:14:30.810 straightforward to swap people's faces. There's a popular YouTube video 0:14:31.170 --> 0:14:35.170 that features tech pioneer Elon Musk's adult face on a 0:14:35.250 --> 0:14:39.290 baby's body, and there's a famous meme where actor Nicholas 0:14:39.330 --> 0:14:43.570 Cage's face replaces those of leading movie actors, both male 0:14:43.650 --> 0:14:47.130 and female. You can put words into people's mouths and 0:14:47.210 --> 0:14:50.970 make them jump and dance and run. You can even 0:14:51.010 --> 0:14:55.290 resurrect powerful figures and have them deliver a fake speech 0:14:55.810 --> 0:15:04.570 about a fake tragedy. From an Altered History, Chapter two, 0:15:05.530 --> 0:15:10.890 Creating Nixon. The text of Nixon's Moon disaster speech that 0:15:10.930 --> 0:15:13.330 we heard at the top of the show is actually 0:15:13.370 --> 0:15:16.330 not fake. As I mentioned, it was written for President 0:15:16.450 --> 0:15:20.690 Nixon as a contingency speech and thankfully never had to 0:15:20.690 --> 0:15:23.970 be delivered. It's an amazing piece of writing. It was 0:15:24.010 --> 0:15:28.130 written by Bill Safire, who was one of Nixon's speech writers. 0:15:28.530 --> 0:15:32.090 This is artist in journalist Francesca Panetta. She's the co 0:15:32.250 --> 0:15:36.850 director of the Nixon Fake or MIT's Moon Disaster Team. 0:15:37.490 --> 0:15:42.410 She's also the creative director in MIT's Center for Advanced Virtuality. 0:15:42.770 --> 0:15:47.330 I was doing experimental journalism at the Guardian newspaper. I 0:15:47.410 --> 0:15:50.490 ran the Guardians Virtual Reality studio for the last three years. 0:15:50.770 --> 0:15:53.250 The second half of the Moon Disaster team is sound 0:15:53.370 --> 0:15:57.650 artist Halsey Bergund. My name is Halsey Bergund. I am 0:15:57.690 --> 0:16:00.690 a sound artist and technologist, and I've had a lot 0:16:00.690 --> 0:16:04.890 of experience with lots of sorts of audio enhanced with technology, 0:16:05.170 --> 0:16:08.530 though this is my first experience with synthetic media, especially 0:16:08.610 --> 0:16:12.090 since I typically focus on authenticity of voice and now 0:16:12.130 --> 0:16:15.570 I'm kind of doing the opposite. So together, Halsey and 0:16:15.650 --> 0:16:19.530 Francesca chose to automate a tragic moment in history that 0:16:19.770 --> 0:16:22.930 never actually happened. I think it all started with it 0:16:23.010 --> 0:16:25.370 being the fiftieth anniversary of the moon landing last year, 0:16:25.490 --> 0:16:28.290 and add on top of that an election cycle in 0:16:28.330 --> 0:16:31.530 this country, and dealing with this information, which is obviously 0:16:32.170 --> 0:16:36.730 very important in election cycles. It was like lightbulbs went 0:16:36.810 --> 0:16:40.250 on and we got very excited about pursuing it. It's 0:16:40.290 --> 0:16:43.810 possible to make mediocre fakes pretty quickly and cheaply, but 0:16:43.930 --> 0:16:47.690 Francesca and Halsey wanted high production values. So how does 0:16:47.730 --> 0:16:51.410 one go about making a first rate fake presidential address? 0:16:51.890 --> 0:16:55.090 There are two components. There's the visuals and there's the audio, 0:16:55.130 --> 0:16:59.370 and they are completely different processes. So we decided to 0:16:59.410 --> 0:17:03.650 go with a video dialogue replacement company called Kenny Ai, 0:17:04.090 --> 0:17:06.650 who would do the visuals for us and then we 0:17:06.730 --> 0:17:10.810 decided to go with re Speech, who are a dialogue 0:17:10.850 --> 0:17:15.690 replacement company for the voice of Nixon. They tackled the 0:17:15.770 --> 0:17:19.090 voice first, the more challenging of the two mediums. What 0:17:19.130 --> 0:17:21.850 we were told to do was to get two to 0:17:21.930 --> 0:17:25.450 three hours worth of Nixon talking. That was pretty easy 0:17:25.450 --> 0:17:28.810 because the Nixon Library has hours and hours of Nixon, 0:17:29.370 --> 0:17:33.690 mainly giving Vietnam's speeches. The Communist armies of North Vietnam 0:17:33.810 --> 0:17:37.290 launched a massive inversion of South Vietnam. That audio was 0:17:37.370 --> 0:17:41.050 then chopped up into chunks between one and three seconds long. 0:17:41.770 --> 0:17:46.610 We found this incredibly patient actor called Lewis D. Wheeler. 0:17:47.250 --> 0:17:51.050 Lewis would listen to the one second clip and then 0:17:51.210 --> 0:17:59.850 he would repeat that and do what I believe was right. 0:18:00.250 --> 0:18:02.410 Re Speech would say to us things like we need 0:18:02.450 --> 0:18:07.370 to change the diagonal attention, which meant nothing to us. Yes, 0:18:07.530 --> 0:18:12.450 we have a whole lot of potential band name going forward. Yeah, 0:18:14.210 --> 0:18:17.410 Synthetic Nixon is another good one. So once we have 0:18:17.810 --> 0:18:21.210 our Nixon model made out of these thousands of tiny clips, 0:18:21.530 --> 0:18:25.890 it means that whatever our actor says will come out 0:18:26.050 --> 0:18:29.450 then in Nixon's voice. So then what we did was 0:18:29.530 --> 0:18:34.130 record the contingency speech of Nixon, and it meant that 0:18:34.170 --> 0:18:39.810 we got Lewis's actually performance but in Nixon's voice. What 0:18:39.850 --> 0:18:42.930 about the video part? I mean, the video was much easier. 0:18:42.970 --> 0:18:45.010 We're talking a couple of days here and a tiny 0:18:45.010 --> 0:18:50.610 amount of data just with Lewis's iPhone. We filmed him 0:18:50.650 --> 0:18:54.370 reading the contingency speech once a couple of minutes of 0:18:54.450 --> 0:18:57.930 him just chatting to camera, and that was it fate 0:18:59.130 --> 0:19:01.650 that the men who went to the Moon to explore 0:19:01.810 --> 0:19:07.770 in peace will stay on. You know. We were told 0:19:07.770 --> 0:19:11.050 by Kenny Ai that everything would be the same in 0:19:11.090 --> 0:19:14.410 the video apart from just the area around the mouth. 0:19:14.850 --> 0:19:18.010 So every gesture of the hand, every blink, every time 0:19:18.050 --> 0:19:20.810 he moved his face, all of that would stay the same, 0:19:21.330 --> 0:19:25.970 but just the mouth basically would change. So we used 0:19:26.210 --> 0:19:30.330 Nixon's resignation speech to have served in this office, it's 0:19:30.450 --> 0:19:35.050 to have felt a very personal sense of it was 0:19:35.370 --> 0:19:38.570 the speech of Nixon that looked the most somber, where 0:19:38.610 --> 0:19:40.810 he seemed to have the most emotion in his face. 0:19:41.490 --> 0:19:45.690 So what actually went on in the computer? Artificial intelligence 0:19:45.970 --> 0:19:50.610 sometimes sounds inscrutable, but the basic ideas are quite simple. 0:19:51.290 --> 0:19:54.290 In this case, it uses a type of computer program 0:19:54.330 --> 0:19:59.010 called an auto encoder. It's trained to take complicated things, 0:19:59.450 --> 0:20:04.010 say spoken sentences or pictures, encode them in a much 0:20:04.050 --> 0:20:08.250 simpler form, and then decode them to recover the original 0:20:08.290 --> 0:20:12.210 as bested. Care the encoder tries to reduce things to 0:20:12.250 --> 0:20:16.490 their essence, throwing away most of the information but keeping 0:20:16.610 --> 0:20:19.690 enough to do a good job of reconstructing it to 0:20:19.690 --> 0:20:23.330 make a deep fake. Here's the trick. Train a speech 0:20:23.370 --> 0:20:27.090 auto encoder for Nixon to Nixon, and a speech auto 0:20:27.210 --> 0:20:31.330 encoder for actor to actor, but force them to use 0:20:31.450 --> 0:20:37.970 the same encoder. Then you can input actor and decoded 0:20:38.330 --> 0:20:41.730 as Nixon. If you have enough data. It's a piece 0:20:41.770 --> 0:20:48.330 of cake around there. Carefully created video, the Moon Disaster 0:20:48.450 --> 0:20:53.170 team created an entire art installation a nineteen sixties living 0:20:53.210 --> 0:20:56.890 room with a fake vintage newspaper sharing the fake tragic 0:20:56.930 --> 0:21:01.730 news while a fake Nixon speaks solemnly on a vintage 0:21:01.770 --> 0:21:05.530 black and white television. Some people, when they were watching 0:21:05.570 --> 0:21:08.570 the installation, they watched a number of times. You'd see them, 0:21:08.610 --> 0:21:10.610 they'd watch at once, then they would watch it again, 0:21:11.410 --> 0:21:13.450 staring at the lips to see if they could see 0:21:13.490 --> 0:21:17.770 any lack of synchronicity. We had some people who thought 0:21:17.810 --> 0:21:21.850 that perhaps Nixon had actually recorded this speech as a 0:21:21.890 --> 0:21:24.810 contingency speech for it to go onto television. Lots of 0:21:24.850 --> 0:21:28.610 folks who were listening, viewing, and even press folks just 0:21:28.690 --> 0:21:30.690 immediately said, oh, the voice is real or whatever you 0:21:30.810 --> 0:21:34.770 said these things that weren't accurate because they just felt 0:21:34.770 --> 0:21:37.330 like there wasn't even a question. I suppose that is 0:21:37.330 --> 0:21:39.490 what we wanted to achieve, But at the same time, 0:21:39.530 --> 0:21:42.530 it was a little bit eye opening and like a 0:21:42.530 --> 0:21:50.890 little scary. You know that that could happen. Chapter three, 0:21:51.610 --> 0:21:55.930 Everybody dance. What do you see as just the wonderful 0:21:57.130 --> 0:22:01.050 upside of having technologies like this? Yeah, I mean a 0:22:01.210 --> 0:22:06.850 aion art is becoming a whole field in itself, so creatively, 0:22:07.170 --> 0:22:11.010 there is enormous potential. One of the potential positive educational 0:22:11.090 --> 0:22:14.730 uses of deep fake technology would be to bring historical 0:22:14.770 --> 0:22:17.930 figures back to life to make learning more durable. I 0:22:17.970 --> 0:22:21.010 think one could do that with bringing Abraham Lincoln back 0:22:21.010 --> 0:22:23.770 to life and having him deliver speeches. Film companies are 0:22:23.770 --> 0:22:26.930 really excited about re enactments. We're already beginning to see 0:22:26.970 --> 0:22:31.010 this in films like Star Wars, when we're bringing people 0:22:31.010 --> 0:22:33.770 like Carrie Fisher back to life. I mean that is 0:22:33.970 --> 0:22:37.090 at the moment not being done through deep fake technologies. 0:22:37.130 --> 0:22:40.650 This is using fatty traditional techniques of CGI at the moment, 0:22:40.890 --> 0:22:43.570 So we still have to see our first deep fake 0:22:43.770 --> 0:22:47.330 big cinema screen release. But this is just to come 0:22:47.450 --> 0:22:50.450 like the technology is getting better and better. Not only 0:22:50.490 --> 0:22:53.530 will we be able to potentially bring back actors and 0:22:53.570 --> 0:22:56.690 actresses who are no longer alive and have them star 0:22:56.770 --> 0:22:59.610 in movies, but an actor could make a model of 0:22:59.650 --> 0:23:02.210 their own voice and then sell the use of that 0:23:02.330 --> 0:23:06.290 voice to anybody to do a voiceover of whatever is wanted, 0:23:06.410 --> 0:23:09.410 and so they could have twenty of the is going 0:23:09.410 --> 0:23:11.410 on at the same time, and the sort of restriction 0:23:11.490 --> 0:23:15.170 of their physical presence is no longer there. And that 0:23:15.250 --> 0:23:17.410 might mean that, you know, Brad Pitt is in everything, 0:23:18.370 --> 0:23:21.530 or it might just mean that lower budget films can 0:23:21.570 --> 0:23:24.050 afford to have some of the higher cost talent. At 0:23:24.090 --> 0:23:26.370 that point, you know, the top twenty actors could just 0:23:26.450 --> 0:23:29.210 do everything. Yes, there's no doubt that there will be 0:23:29.210 --> 0:23:33.050 winners and losers from these technologies, but the potential of 0:23:33.090 --> 0:23:36.530 synthetic media goes way beyond the arts. There are possible 0:23:36.610 --> 0:23:40.730 medical and therapeutic applications. There are companies that are working 0:23:40.810 --> 0:23:44.210 very hard to allow people who have either lost their 0:23:44.290 --> 0:23:46.330 voice or who never had a voice, to be able 0:23:46.370 --> 0:23:49.890 to speak in a way that is either how they 0:23:49.970 --> 0:23:52.690 used to speak or in a way that isn't a 0:23:52.690 --> 0:23:56.570 canned voice that everybody has. Alexei ePROs and his students 0:23:56.690 --> 0:24:01.410 discovered potential uses of synthetic media and medicine quite unintentionally 0:24:01.810 --> 0:24:06.050 while working on their Everybody Dance Now project that could 0:24:06.050 --> 0:24:10.970 turn anyone into a ballerina. Were kind of surprised for 0:24:11.210 --> 0:24:14.650 all the positive feedback we got. We've got emails from 0:24:14.770 --> 0:24:17.650 people who were quadriplegic and they asked us if we 0:24:17.690 --> 0:24:21.450 could make them dance, and it was very unexpected. So 0:24:21.490 --> 0:24:24.570 now we are trying to get the software to be 0:24:24.650 --> 0:24:27.890 in a state where people can use it, because yeah, 0:24:27.930 --> 0:24:33.530 it's somehow it did hit a nerve with folks. Chapter 0:24:33.690 --> 0:24:39.250 four Unicorns in the Andes. The past few years have 0:24:39.370 --> 0:24:43.250 seen amazing advances in the creation of synthetic media through 0:24:43.330 --> 0:24:47.690 artificial intelligence. The technology now goes far beyond fitting one 0:24:47.730 --> 0:24:51.610 face over another face in a video. A recent breakthrough 0:24:51.690 --> 0:24:56.130 has made it possible to create entirely new and very 0:24:56.250 --> 0:25:01.370 convincing content out of thin air. The breakthrough called generative 0:25:01.530 --> 0:25:06.690 adversarial networks or GAMS, came from a machine learning researcher 0:25:06.730 --> 0:25:11.530 at Google named Ian Goodfellow. Like auto encoders, the basic 0:25:11.610 --> 0:25:15.930 idea is simple but brilliant. Suppose you want to create 0:25:16.050 --> 0:25:20.570 amazingly realistic photos of people who don't exist. While you 0:25:20.610 --> 0:25:24.810 build a GAN consisting of two computer programs, a photo 0:25:25.010 --> 0:25:29.170 generator that learns to generate fake photos and a photo 0:25:29.330 --> 0:25:35.170 discriminator that learns to discriminate or identify fake photos from 0:25:35.170 --> 0:25:38.850 a vast collection of real photos. You then let the 0:25:38.850 --> 0:25:43.690 two programs compete, continually tweaking their code to outsmart each other. 0:25:44.490 --> 0:25:48.330 By the time they're done, the GAN can generate amazingly 0:25:48.330 --> 0:25:51.730 convincing fakes. You can see for yourself if you go 0:25:51.770 --> 0:25:55.810 to the website this Person does Not Exist dot com. 0:25:56.650 --> 0:25:59.890 Every time you refresh the page, you're shown a new 0:26:00.130 --> 0:26:04.410 uncanny image of a person who, as the website says, 0:26:04.890 --> 0:26:09.450 does not and never did exist. Francescan I actually tried 0:26:09.530 --> 0:26:17.650 out the website. This young Asian woman. She's got great complexion. 0:26:17.890 --> 0:26:20.850 Envious of that neat black hair with a fringe pink 0:26:20.850 --> 0:26:24.130 lipstick and a slightly dreamy look as she's kind of 0:26:24.130 --> 0:26:30.330 gazing off to her left. Oh, here's a woman who 0:26:30.330 --> 0:26:32.770 looks like she could be a neighbor of mine in Cambridge, 0:26:33.410 --> 0:26:38.090 probably about sixty five. She's got nice wire framed glasses, 0:26:38.370 --> 0:26:43.610 layered hair. Her earrings don't actually match, but that could 0:26:43.610 --> 0:26:46.690 just be her distinctive style. I mean, of course, she 0:26:47.050 --> 0:26:52.330 doesn't really exist. It's hard to argue that gams aren't 0:26:52.330 --> 0:26:57.970 creating original art. In fact, an artist collective recently used 0:26:58.010 --> 0:27:03.050 a GAM to create a French Impressionist style portrait. When 0:27:03.170 --> 0:27:07.210 Christie's sold it at auction, it fetched an eye popping 0:27:07.290 --> 0:27:11.970 four hundred and thirty two thousand dollars. Alexei Efros, the 0:27:12.010 --> 0:27:17.130 Berkeley professor, recently pushed gans a step further, creating something 0:27:17.210 --> 0:27:21.770 called cycle gans. By connecting two gans together in a 0:27:21.770 --> 0:27:27.290 clever way, cycle gans can transform a monet painting into 0:27:27.290 --> 0:27:31.210 what's seemingly a photograph of the same scene, or turn 0:27:31.250 --> 0:27:34.930 a summer landscape into a winter landscape of the same view. 0:27:35.810 --> 0:27:39.290 Alexei's cycle gans seem like magic. If you were to 0:27:39.290 --> 0:27:44.810 add in virtual reality, the possibilities become mind blowing. You 0:27:45.090 --> 0:27:51.490 may be reminiscing about walking down Saint German and Paris 0:27:51.570 --> 0:27:54.450 and with a few clicks. You are there, and you're 0:27:54.530 --> 0:27:57.410 walking down the boulevard, and you're looking at all the buildings, 0:27:57.450 --> 0:28:00.970 and maybe you can even switch to a different year. 0:28:01.130 --> 0:28:06.330 And I think that is I think very exciting as 0:28:06.370 --> 0:28:10.490 a way to mentally travel to different places. So if 0:28:10.490 --> 0:28:12.650 you do this in VR, I mean, can you imagine 0:28:13.010 --> 0:28:16.810 classes going on a class visit to ancient Rome. That's right, 0:28:17.170 --> 0:28:21.890 you could imagine from how a particular city like Chrome. 0:28:22.290 --> 0:28:24.770 Luke's now trying to extrapolate to how it looked in 0:28:24.810 --> 0:28:29.210 the past. It turns out that gans aren't just transforming images. 0:28:29.890 --> 0:28:32.690 I spoke with a friend who's very familiar with another 0:28:32.810 --> 0:28:37.250 remarkable application of the technology. My name is Reid Hoffman. 0:28:37.370 --> 0:28:40.650 I'm a podcaster of Master's Scale. I'm a partner at Greylock, 0:28:40.690 --> 0:28:43.570